skillgram
Skill governance for AI agents. One library, a contract that gates it, every agent reading the same thing.
skillgram is a registry of SKILL.md files that lives in one directory on
disk, plus the layer that governs it: a contract every skill has to pass, provenance for each one,
declared usage, and a ledger of what was propagated where. Claude Code, Codex, Cursor, Hermes,
Reasonix, Gemini CLI, OpenCode, Kiro, Qwen Code, Aider and Copilot all read the same library.
Links are the mechanism; governance is the product. No LLM in the loop unless you ask for one, and no claim that it can see what an agent reads.
Install
pip install skillgram
Or from source, and then the whole machine in one idempotent command:
git clone https://github.com/Xhuk/skillgram
cd skillgram
python skillgram.py install --dry-run # the exact commands, nothing written
python skillgram.py install # harness -> hub -> MCP -> project scope
Python 3.9+, standard library only. Junctions on Windows, symlinks on macOS/Linux.
The model
~/.agents/skills/<category>/<skill>/SKILL.md is the only real copy.lint --contract: trigger, hard rules, decisions, steps, output. Or it does not ship.SKILLMETA.json: who wrote it, when, from where, and the hash that makes updates possible.skills/ folder.use + mount give a repo exactly its set, not the whole library.skillgram mcp exposes 7 tools: search, missing, use, status, propose, touch, list.touch + list --used: what an agent says it used, per project. It never claims to see what it reads.What it is not
Not a prompt manager, not a memory store, not an embeddings index. It links and administers skills your agents already read, and it never guesses which harness is asking.